Forced alignment by Kaldi is added.

This commit is contained in:
yemaozi88 2018-08-20 22:50:53 +02:00
parent d56ef7f075
commit 22b9ae966b
7 changed files with 451 additions and 161 deletions

Binary file not shown.

View File

@ -15,6 +15,7 @@ Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Solution Items", "Solution
..\forced_alignment\forced_alignment\htk_dict.py = ..\forced_alignment\forced_alignment\htk_dict.py
..\forced_alignment\forced_alignment\lexicon.py = ..\forced_alignment\forced_alignment\lexicon.py
..\forced_alignment\forced_alignment\mlf.py = ..\forced_alignment\forced_alignment\mlf.py
..\accent_classification\accent_classification\output_confusion_matrix.py = ..\accent_classification\accent_classification\output_confusion_matrix.py
..\forced_alignment\forced_alignment\pronunciations.py = ..\forced_alignment\forced_alignment\pronunciations.py
..\forced_alignment\forced_alignment\pyhtk.py = ..\forced_alignment\forced_alignment\pyhtk.py
..\forced_alignment\forced_alignment\scripts.py = ..\forced_alignment\forced_alignment\scripts.py

View File

@ -22,12 +22,11 @@ dataset_list = ['devel', 'test', 'train']
extract_features = 0
make_feature_list = 0
conv_lexicon = 0
check_lexicon = 1
check_lexicon = 0
make_mlf = 0
combine_files = 0
flat_start = 0
train_model = 0
forced_alignment = 0
train_model = 1
sys.path.append(os.path.join(os.path.dirname(sys.path[0]), curr_dir))
@ -288,7 +287,7 @@ if flat_start:
## ======================= estimate monophones =======================
if train_model:
iter_num_max = 3
for mix_num in [16, 32, 64, 128]:
for mix_num in [128, 256, 512, 1024]:
for iter_num in range(1, iter_num_max+1):
print("===== mix{}, iter{} =====".format(mix_num, iter_num))
iter_num_pre = iter_num - 1
@ -315,5 +314,6 @@ if train_model:
fout.write("MU %d {*.state[2-4].mix}" % (mix_num_next))
subprocessStr = 'HHEd -T 1 -H ' + modelN_dir + '\\' + hmmdefs_name + ' -M ' + modelN_dir_next + ' ' + header_file + ' ' + phonelist
subprocess.call(subprocessStr, shell=True)

View File

@ -31,6 +31,9 @@
<Compile Include="performance_check.py">
<SubType>Code</SubType>
</Compile>
<Compile Include="pyKaldi.py">
<SubType>Code</SubType>
</Compile>
</ItemGroup>
<ItemGroup>
<Content Include="config.ini" />

View File

@ -2,4 +2,4 @@
config_hcopy = c:\cygwin64\home\Aki\acoustic_model\config\config.HCopy
config_train = c:\cygwin64\home\Aki\acoustic_model\config\config.train
mkhmmdefs_pl = c:\cygwin64\home\Aki\acoustic_model\src\acoustic_model\mkhmmdefs.pl
FAME_dir = c:\OneDrive\Research\rug\experiments\friesian\corpus
FAME_dir = C:\OneDrive\WSL\kaldi-trunk\egs\fame\s5\corpus

View File

@ -4,10 +4,12 @@ import csv
import subprocess
import configparser
from collections import Counter
import re
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.metrics import confusion_matrix
## ======================= functions =======================
@ -30,26 +32,64 @@ def read_fileFA(fileFA):
return ' '.join(phones)
#####################
## USER DEFINE ##
#####################
def make_dic(word, pronvar_, fileDic, output_type):
"""
make dict files which can be used for HTK.
param word: target word.
param pronvar_: pronunciation variant. nx2 (WORD /t pronunciation) ndarray.
param fileDic: output dic file.
param output_type: 0:full, 1:statistics, 2:frequency <2% entries are removed. 3:top 3.
"""
#assert(output_type < 4 and output_type >= 0, 'output_type should be an integer between 0 and 3.')
if output_type == 0: # full
pronvar = np.unique(pronvar_)
with open(fileDic, 'w') as f:
for pvar in pronvar:
f.write('{0}\t{1}\n'.format(WORD, pvar))
else:
c = Counter(pronvar_)
total_num = sum(c.values())
with open(fileDic, 'w') as f:
if output_type == 3:
for key, value in c.most_common(3):
f.write('{0}\t{1}\n'.format(WORD, key))
else:
for key, value in c.items():
percentage = value/total_num*100
if output_type == 1: # all
f.write('{0}\t{1:.2f}\t{2}\t{3}\n'.format(value, percentage, WORD, key))
elif output_type == 2: # less than 2 percent
if percentage < 2:
f.write('{0}\t{1}\n'.format(WORD, key))
## ======================= user define =======================
curr_dir = r'C:\Users\Aki\source\repos\acoustic_model\acoustic_model'
config_ini = curr_dir + '\\config.ini'
forced_alignment_module = r'C:\Users\Aki\source\repos\forced_alignment'
forced_alignment_module_old = r'C:\OneDrive\Research\rug\code\forced_alignment\forced_alignment'
ipa_xsampa_converter_dir = r'C:\Users\Aki\source\repos\ipa-xsama-converter'
accent_classification_dir = r'C:\Users\Aki\source\repos\accent_classification\accent_classification'
csvfile = r"C:\OneDrive\Research\rug\stimmen\Frisian Variants Picture Task Stimmen.csv"
experiments_dir = r'C:\OneDrive\Research\rug\experiments'
data_dir = experiments_dir + '\\stimmen\\data'
csvfile = data_dir + '\\Frisian Variants Picture Task Stimmen.csv'
cygwin_dir = r'C:\cygwin64\home\Aki\acoustic_model'
# procedure
convert_phones = 0
make_dic_files = 0
make_dic_files_short = 0
do_forced_alignment = 0
eval_forced_alignment = 1
do_forced_alignment_htk = 0
make_kaldi_data_files = 0
make_kaldi_lexicon_txt = 0
load_forced_alignment_kaldi = 1
eval_forced_alignment = 0
@ -67,6 +107,10 @@ import acoustic_model_functions as am_func
sys.path.append(forced_alignment_module_old)
import pyHTK
# to output confusion matrix
sys.path.append(accent_classification_dir)
from output_confusion_matrix import plot_confusion_matrix
## ======================= load variables =======================
config = configparser.ConfigParser()
@ -146,53 +190,37 @@ word_list = np.unique(words)
## ======================= make dict files used for HTK. ======================
if make_dic_files:
output_dir = experiments_dir + r'\stimmen\dic'
output_type = 2
output_dir = experiments_dir + r'\stimmen\dic_short'
for word in word_list:
WORD = word.upper()
fileDic = output_dir + '\\' + word + '.dic'
# make dic file.
# pronunciation variant of the target word.
pronvar_ = pronunciations[words == word]
pronvar = np.unique(pronvar_)
# remove ''
pronvar_ = np.delete(pronvar_, np.where(pronvar_==''))
with open(fileDic, 'w') as f:
for pvar in pronvar:
f.write('{0}\t{1}\n'.format(WORD, pvar))
# make dic file.
make_dic(word, pronvar_, fileDic, output_type)
## ======================= make dict files for most popular words. ======================
if make_dic_files_short:
output_dir = experiments_dir + r'\stimmen\dic'
#word = word_list[3]
for word in word_list:
WORD = word.upper()
fileStat = output_dir + '\\' + word + '_stat.csv'
pronvar = pronunciations[words == word]
c = Counter(pronvar)
total_num = sum(c.values())
with open(fileStat, 'w') as f:
for key, value in c.items():
f.write('{0}\t{1:.2f}\t{2}\t{3}\n'.format(value, value/total_num*100, WORD, key))
## ======================= forced alignment =======================
if do_forced_alignment:
## ======================= forced alignment using HTK =======================
if do_forced_alignment_htk:
configHVite = cygwin_dir + r'\config\config.HVite'
filePhoneList = experiments_dir + r'\friesian\acoustic_model\config\phonelist_friesian.txt'
wav_dir = experiments_dir + r'\stimmen\wav'
#for hmm_num in [1, 2, 4, 8, 16, 32, 64, 128]:
for hmm_num in [64]:
#hmm_num = 128
for hmm_num in [1, 2, 4, 8, 16, 32, 64, 128, 256]:
hmm_num_str = str(hmm_num)
AcousticModel = experiments_dir + r'\friesian\acoustic_model\model\hmm' + hmm_num_str + r'-3\hmmdefs'
AcousticModel = experiments_dir + r'\friesian\acoustic_model\model\hmm' + hmm_num_str + r'-2\hmmdefs'
predictions = []
file_num_max = len(filenames)
for i in range(0, file_num_max):
#for i in range(500, 502):
print('=== {0}/{1} ==='.format(i, file_num_max))
filename = filenames[i]
fileWav = wav_dir + '\\' + filename
@ -206,8 +234,8 @@ if do_forced_alignment:
with open(fileLab, 'w') as f:
lines = f.write(WORD)
fileDic = experiments_dir + r'\stimmen\dic_short' + '\\' + word + '.dic'
fileFA = experiments_dir + r'\stimmen\FA_short' + '\\' + filename.replace('.wav', '.txt') + hmm_num_str
fileDic = experiments_dir + r'\stimmen\dic_top3' + '\\' + word + '.dic'
fileFA = experiments_dir + r'\stimmen\FA' + '\\' + filename.replace('.wav', '.txt') + hmm_num_str
pyHTK.doHVite(fileWav, fileLab, fileDic, fileFA, configHVite, filePhoneList, AcousticModel)
prediction = read_fileFA(fileFA)
@ -224,11 +252,220 @@ if do_forced_alignment:
np.save(data_dir + '\\match_hmm' + hmm_num_str + '.npy', match)
## ======================= make files which is used for forced alignment by Kaldi =======================
if make_kaldi_data_files:
wav_dir = r'c:\OneDrive\WSL\kaldi-trunk\egs\fame\s5\corpus\stimmen'
kaldi_work_dir = r'C:\OneDrive\WSL\kaldi-trunk\egs\fame\s5'
kaldi_data_dir = os.path.join(kaldi_work_dir, 'data', 'alignme')
kaldi_dict_dir = os.path.join(kaldi_work_dir, 'data', 'local', 'dict')
htk_dict_dir = os.path.join(experiments_dir, 'stimmen', 'dic_top3')
wav_scp = os.path.join(kaldi_data_dir, 'wav.scp')
text_file = os.path.join(kaldi_data_dir, 'text')
utt2spk = os.path.join(kaldi_data_dir, 'utt2spk')
lexicon_txt = os.path.join(kaldi_dict_dir, 'lexicon.txt')
predictions = []
file_num_max = len(filenames)
# remove previous files.
if os.path.exists(wav_scp):
os.remove(wav_scp)
if os.path.exists(text_file):
os.remove(text_file)
if os.path.exists(utt2spk):
os.remove(utt2spk)
f_wav_scp = open(wav_scp, 'a', encoding="utf-8", newline='\n')
f_text_file = open(text_file, 'a', encoding="utf-8", newline='\n')
f_utt2spk = open(utt2spk, 'a', encoding="utf-8", newline='\n')
# make wav.scp, text, and utt2spk files.
for i in range(0, file_num_max):
#for i in range(400, 410):
print('=== {0}/{1} ==='.format(i+1, file_num_max))
filename = filenames[i]
wav_file = wav_dir + '\\' + filename
if os.path.exists(wav_file):
speaker_id = 'speaker_' + str(i).zfill(4)
utterance_id = filename.replace('.wav', '')
utterance_id = utterance_id.replace(' ', '_')
utterance_id = speaker_id + '-' + utterance_id
# wav.scp file
wav_file_unix = wav_file.replace('\\', '/')
wav_file_unix = wav_file_unix.replace('c:/', '/mnt/c/')
f_wav_scp.write('{0} {1}\n'.format(utterance_id, wav_file_unix))
# text file
word = words[i].lower()
f_text_file.write('{0}\t{1}\n'.format(utterance_id, word))
# utt2spk
f_utt2spk.write('{0} {1}\n'.format(utterance_id, speaker_id))
f_wav_scp.close()
f_text_file.close()
f_utt2spk.close()
## ======================= make lexicon txt which is used by Kaldi =======================
if make_kaldi_lexicon_txt:
kaldi_work_dir = r'C:\OneDrive\WSL\kaldi-trunk\egs\fame\s5'
kaldi_dict_dir = os.path.join(kaldi_work_dir, 'data', 'local', 'dict')
lexicon_txt = os.path.join(kaldi_dict_dir, 'lexicon.txt')
option_num = 5
# remove previous file.
if os.path.exists(lexicon_txt):
os.remove(lexicon_txt)
mapping = convert_xsampa2ipa.load_converter('xsampa', 'ipa', ipa_xsampa_converter_dir)
with open(csvfile, encoding="utf-8") as fin:
lines = csv.reader(fin, delimiter=';', lineterminator="\n", skipinitialspace=True)
next(lines, None) # skip the headers
filenames = []
words = []
pronunciations = []
p = []
for line in lines:
if line[1] is not '' and len(line) > 5:
filenames.append(line[0])
words.append(line[1])
pron_xsampa = line[3]
pron_ipa = convert_xsampa2ipa.conversion('xsampa', 'ipa', mapping, pron_xsampa)
pron_ipa = pron_ipa.replace('ː', ':')
# adjust to phones used in the acoustic model.
pronunciations.append(pron_ipa)
# check if all phones are in the phonelist of the acoustic model.
#'y', 'b', 'ɾ', 'u', 'ɔ:', 'ø', 't', 'œ', 'n', 'ɒ', 'ɐ', 'f', 'o', 'k', 'x', 'ɡ', 'v', 's', 'ɛ:', 'ɪ:', 'ɑ', 'ɛ', 'a', 'd', 'z', 'ɪ', 'ɔ', 'l', 'i:', 'm', 'p', 'a:', 'i', 'e', 'j', 'o:', 'ʁ', 'h', ':', 'e:', 'ə', 'æ', 'χ', 'w', 'r', 'ə:', 'sp', 'ʊ', 'u:', 'ŋ'
filenames = np.array(filenames)
words = np.array(words)
wordlist = np.unique(words)
pronunciations = np.array(pronunciations)
# output lexicon.txt
#f_lexicon_txt = open(lexicon_txt, 'a', encoding="utf-8", newline='\n')
pronvar_list_all = []
for word in word_list:
# pronunciation variant of the target word.
pronvar_ = pronunciations[words == word]
# remove ''
pronvar_ = np.delete(pronvar_, np.where(pronvar_==''))
c = Counter(pronvar_)
total_num = sum(c.values())
for key, value in c.most_common(option_num):
#print('{0}\t{1}\t{2}\t{3}'.format(word, key, value, total_num))
key = key.replace('æ', 'ɛ')
key = key.replace('ɐ', 'a')
key = key.replace('ɑ', 'a')
key = key.replace('ɾ', 'r')
key = key.replace('ʁ', 'r')
key = key.replace('ʊ', 'u')
key = key.replace('χ', 'x')
#print('-->{0}\t{1}\t{2}\t{3}\n'.format(word, key, value, total_num))
# make possible pronounciation variant list.
pronvar_list = [key]
while 'ø:' in ' '.join(pronvar_list) or 'œ' in ' '.join(pronvar_list) or 'ɒ' in ' '.join(pronvar_list):
pronvar_list_ = []
for p in pronvar_list:
if 'ø:' in p:
pronvar_list_.append(p.replace('ø:', 'ö'))
pronvar_list_.append(p.replace('ø:', 'ö:'))
if 'œ' in p:
pronvar_list_.append(p.replace('œ', 'ɔ̈'))
pronvar_list_.append(p.replace('œ', 'ɔ̈:'))
if 'ɒ' in p:
pronvar_list_.append(p.replace('ɒ', 'ɔ̈'))
pronvar_list_.append(p.replace('ɒ', 'ɔ̈:'))
pronvar_list = np.unique(pronvar_list_)
for pronvar_ in pronvar_list:
split_ipa = convert_phone_set.split_ipa_fame(pronvar_)
pronvar_out = ' '.join(split_ipa)
pronvar_list_all.append([word, pronvar_out])
# output
pronvar_list_all = np.array(pronvar_list_all)
pronvar_list_all = np.unique(pronvar_list_all, axis=0)
#f_lexicon_txt.write('<UNK>\tSPN\n')
#for line in pronvar_list_all:
# f_lexicon_txt.write('{0}\t{1}\n'.format(line[0].lower(), line[1]))
#f_lexicon_txt.close()
## ======================= load kaldi forced alignment result =======================
if load_forced_alignment_kaldi:
kaldi_work_dir = r'C:\OneDrive\WSL\kaldi-trunk\egs\fame\s5'
phones_txt = kaldi_work_dir + '\\data\\lang\\phones.txt'
merged_alignment_txt = kaldi_work_dir + '\\exp\\tri1_alignme\\merged_alignment.txt'
filenames = np.load(data_dir + '\\filenames.npy')
words = np.load(data_dir + '\\words.npy')
pronunciations = np.load(data_dir + '\\pronunciations_ipa.npy')
pronvar_list_all = np.load(data_dir + '\\pronvar_list_all.npy')
word_list = np.unique(words)
# load the mapping between phones and ids.
with open(phones_txt, 'r', encoding="utf-8") as f:
mappings = f.read().split('\n')
phones = []
phone_ids = []
for m in mappings:
m = m.split(' ')
if len(m) > 1:
phones.append(m[0])
phone_ids.append(int(m[1]))
with open(merged_alignment_txt, 'r') as f:
lines = f.read()
lines = lines.split('\n')
fa_filenames = []
fa_pronunciations = []
filename_ = ''
pron = []
for line in lines:
line = line.split(' ')
if len(line) == 5:
filename = line[0]
if filename == filename_:
phone_id = int(line[4])
#if not phone_id == 1:
phone = phones[phone_ids.index(phone_id)]
pron_ = re.sub(r'_[A-Z]', '', phone)
if not pron_ == 'SIL':
pron.append(pron_)
else:
fa_filenames.append(re.sub(r'speaker_[0-9]{4}-', '', filename))
fa_pronunciations.append(' '.join(pron))
pron = []
filename_ = filename
# correct or not.
for filename, fa_pronunciation in zip(fa_filenames, fa_pronunciations):
## ======================= evaluate the result of forced alignment =======================
if eval_forced_alignment:
#for hmm_num in [1, 2, 4, 8, 16, 32, 64]:
hmm_num = 64
match_num = []
for hmm_num in [1, 2, 4, 8, 16, 32, 64, 128, 256]:
#hmm_num = 256
hmm_num_str = str(hmm_num)
match = np.load(data_dir + '\\match_hmm' + hmm_num_str + '.npy')
@ -236,9 +473,10 @@ if eval_forced_alignment:
if 1:
pronunciation_variants = np.array(['WORD', 'pronunciation']).reshape(1, 2)
for word in word_list:
fileDic = experiments_dir + r'\stimmen\dic_short' + '\\' + word + '.dic'
fileDic = experiments_dir + r'\stimmen\dic_top3' + '\\' + word + '.dic'
pronunciation_variants = np.r_[pronunciation_variants, pyHTK.loadHTKdic(fileDic)]
# see only words which appears in top 3.
match_short = []
for line in match:
word = line[0]
@ -254,4 +492,26 @@ if eval_forced_alignment:
# number of match
total_match = sum(match[:, 1] == match[:, 2])
print("{}: {}/{}".format(hmm_num_str, total_match, match.shape[0]))
match_num.append([hmm_num, total_match, match.shape[0]])
# number of mixtures vs accuracy
match_num = np.array(match_num)
plt.xscale("log")
plt.plot(match_num[:, 0], match_num[:, 1]/match_num[0, 2], 'o-')
plt.xlabel('number of mixtures', fontsize=14, fontweight='bold')
plt.ylabel('accuracy', fontsize=14, fontweight='bold')
plt.show()
# confusion matrix
#dir_out = r'C:\OneDrive\Research\rug\experiments\stimmen\result'
#word_list = np.unique(match[:, 0])
#for word in word_list:
# match_ = match[match[:, 0] == word, :]
# cm = confusion_matrix(match_[:, 1], match_[:, 2])
# pronvar = pronunciation_variants[pronunciation_variants[:, 0] == word.upper(), 1]
# plt.figure()
# plot_confusion_matrix(cm, classes=pronvar, normalize=True)
# plt.savefig(dir_out + '\\cm_' + word + '.png')

26
acoustic_model/pyKaldi.py Normal file
View File

@ -0,0 +1,26 @@
import os
import sys
forced_alignment_module = r'C:\Users\Aki\source\repos\forced_alignment'
## ======================= add paths =======================
sys.path.append(forced_alignment_module)
from forced_alignment import convert_phone_set
htk_dict_file = r'C:\OneDrive\Research\rug\experiments\stimmen\dic_top3\Reus.dic'
#kaldi_lexicon = r'C:\OneDrive\WSL\kaldi-trunk\egs\fame\s5\data\lang\phones\'
alignment_txt = r'C:\OneDrive\WSL\kaldi-trunk\egs\fame\s5\exp\tri1_alignme\merged_alignment.txt'
phones_txt = r'C:\OneDrive\WSL\kaldi-trunk\egs\fame\s5\exp\tri1_alignme\phones.txt'
phone_map_txt = r'C:\OneDrive\WSL\kaldi-trunk\egs\fame\s5\data\local\lang\phone_map.txt'
with open(phone_map_txt, 'r', encoding="utf-8") as f:
lines = f.read()
lines = lines.split('\n')
with open(alignment_txt, 'r', encoding="utf-8") as f:
lines =
#phone_in = [line for line in lines if 'SIL' in line]
#if len(phone_in) == 1: